Can't figure out how to cull low-scoring skeletons

[olinesn]

Hello,

I've created a rich dataset where I'd like to cull low-scoring skeletons (below 0.7) during inference before tracking. As an example, if I have a video with two mice, and SLEAP provides two skeletons scoring 0.4 and 0.9, I'd like the 0.4 skeleton to be deleted before tracking.

Perhaps I'm doing something wrong, but I can't figure out how to delete these skeletons aside from manually in the GUI ('Labels > Delete All Predictions with Low Score'). Here's my syntax:
sleap-track --verbosity rich --tracking.tracker flow --tracking.target_instance_count 2 --tracking.pre_cull_to_target 2 --tracking.pre_cull_iou_threshold .9 --tracking.post_connect_single_breaks 1 --tracking.clean_instance_count 2 --tracking.clean_iou_threshold .9 --tracking.similarity instance --tracking.track_window 10 --model "..." --model "..." "....mp4"

The result includes two tracks, which both have skeletons below 0.9. Is it possible that the above syntax culls skeletons until there are only two left, and then leaves them regardless of their score?

Thanks!

[roomrys]

Hi @olinesn,

There is an indirect way of filtering instances with a low score using the --peak_threshold argument (default: 0.2). This will determine whether to consider a proposed body part location valid based on the node score (not the score of the entire instance - which is the average of all the node scores). Setting a higher peak_threshold will yield less predicted nodes, but the nodes that are returned will have a higher score. We should just have a command line option to delete the instances based on score (i.e. what the user sees).

The pre_cull_iou_threshold and clean_iou_threshold arguments cull instances based on the intersection over union of the instances' bounding boxes (not their scores). Also, if you pre_cull_iou_threshold before tracking, then you do not need to clean_iou_threshold after tracking as the instances that meet this criteria will have already been culled.

Let us know if setting the peak_threshold helps (although it beats around the bush a bit).

Thanks,
Liezl

[olinesn]

Hi Liezl,

This is huge, I prefer your suggestion to my proposed solution, that way the good nodes from the bad skeletons stay, and I don't need a skeleton threshold at all!

Thanks also for the tip about the cleanup step, good to know it's redundant.